AI in Clinical Practice: Working Smarter Without Losing What Matters

The Clinical Case for AI: Why This Matters Now

Clinician burnout is not an abstract concern. A 2024 report from the American Medical Association found that nearly half of all physicians reported at least one symptom of burnout, with administrative burden — particularly documentation — consistently cited as a primary driver.¹ For functional and lifestyle medicine practitioners, whose consultations tend to be longer, more complex, and more narrative in nature than conventional appointments, the documentation burden is often acute.

Artificial intelligence tools are entering clinical workflows at precisely this pressure point. The questions for practitioners are practical ones: which tools actually work, which have genuine evidence behind them, what are the real limitations, and what will this cost? This review addresses each of these questions directly, drawing on the most current peer-reviewed literature and independent comparative analyses available as of 2026.

The field divides into two broad areas of clinical application: AI-assisted documentation (ambient scribes and transcription tools) and AI-assisted clinical decision support (diagnostic reasoning, risk stratification, and laboratory interpretation). The evidence base for each is meaningfully different, and practitioners should understand that distinction before adopting any tool.

AI Documentation Tools: What the Trials Show

The strongest and most consistent evidence in clinical AI currently concerns ambient documentation — tools that listen to a practitioner-patient conversation and automatically generate structured clinical notes. The technology has matured significantly since its early iterations, and as of 2025, several large-scale trials have produced credible data.

A three-arm randomised controlled trial conducted across a large academic health system in California (Lukac et al., published as a preprint on medRxiv, 2025) enrolled 238 outpatient physicians across 14 specialties and compared Microsoft DAX Copilot and Nabla against standard care over two months. Both AI scribe arms produced statistically significant reductions in documentation time and decreases in reported burnout indicators. The Nabla arm showed slightly stronger effects on burnout metrics, while DAX showed marginally greater time savings for high-frequency users. Critically, benefits were observed across all 14 specialties tested, suggesting these tools are not specialty-specific in their utility.²

A narrative review of 18 studies published in PMC (2025) found consistent documentation time reductions across all major platforms. Nuance DAX Copilot reduced note-writing time by 20.4% in one quality improvement study (Duggan et al., from 10.3 to 8.2 minutes per note, p<0.001). The Abridge platform produced a reduction from 6.2 to 5.3 minutes per encounter (Stults et al., 2025, p<0.001). Nabla generated clinical notes in under 20 seconds in operational testing.³

Beyond time savings, several studies measured the effect on clinician wellbeing. Shah et al. found that physicians using DAX Copilot showed a statistically significant drop in work exhaustion scores on the Stanford Professional Fulfillment Index (decrease of 1.94 points, p<0.001). Misurac et al. reported burnout scores declining from 4.16 to 3.16 (p=0.005) in Nabla users.³ These are not trivial effects — they represent a meaningful shift in a problem that functional medicine practitioners face acutely.

Leading AI Documentation Platforms: A Comparative Assessment

AI for Diagnostic Support and Clinical Decision-Making

The evidence landscape here is more nuanced and requires careful interpretation. In controlled research settings, large language models have demonstrated impressive performance on diagnostic reasoning tasks — several 2025 studies showed AI systems matching or outperforming physicians on standardised diagnostic benchmarks (Brodeur et al., 2025; Buckley et al., 2025, cited in the Stanford-Harvard State of Clinical AI report).⁴ However, the same report notes a consistent and important finding: when models were required to ask follow-up questions, manage incomplete information, or revise their reasoning as new clinical details emerged, performance declined substantially.

On tests designed to measure diagnostic reasoning under uncertainty — the everyday condition of functional medicine — AI systems performed closer to medical students than to experienced clinicians, and showed a tendency to commit strongly to a diagnosis even in ambiguous presentations.⁴ This overconfidence is not a trivial problem in clinical contexts where uncertainty is the norm rather than the exception.

A 2025 systematic review in Frontiers in Digital Health confirmed that randomised controlled trials represent only a minority of primary studies in clinical AI, with the majority relying on retrospective analysis of fixed case datasets — conditions far removed from live clinical encounters.⁵ Regulatory approvals for AI diagnostic tools remain concentrated in radiology and cardiology, where image-based pattern recognition is well-suited to AI's strengths.

Laboratory and Testing Integration

One of the more practically significant applications for functional medicine is AI-assisted interpretation of complex laboratory panels. Several platforms now offer integration with functional testing providers — including Dutch hormone panels, organic acids, comprehensive stool analysis, and nutrigenomic testing — to generate interpretive summaries and pattern recognition across multiple analytes simultaneously.

Platforms worth awareness in this space include Rupa Health (a testing aggregation platform with AI-assisted reporting), Fullscript's AI recommendation engine (for supplementation protocols informed by lab data), and Everly Health's integrated lab ordering and interpretation tools. These are not diagnostic AI in the classical sense but represent a meaningful clinical workflow improvement for practitioners managing high-volume complex testing.

The evidence base for these functional medicine-specific integrations remains largely observational and vendor-reported. Independent validation is limited, and practitioners should apply the same critical lens to AI-generated functional lab interpretations as they would to any clinical tool without peer-reviewed validation.

AI-Assisted Patient Communication and Intake

A growing body of tools assists with patient-facing communication: automated intake questionnaires, symptom trackers, and between-appointment messaging. Notable examples include Elation Health (an EHR with built-in AI triage), Healthie (a functional medicine-oriented practice management platform with AI-assisted client intake and messaging), and Osmind (primarily for mental health but expanding into integrative medicine).

A 2025 UK survey of general practitioners (PMC, n=1005) found that among those using generative AI in clinical practice, documentation (29%) and generating differential diagnoses (28%) were the most common applications. The rate of AI adoption among GPs had risen significantly from 20% in early 2024 — indicating rapid uptake in primary care contexts similar to functional medicine practice.⁶

The Hallucination Problem: A Non-Negotiable Caution

Any honest review of clinical AI must address the hallucination problem directly. AI language models — including those powering documentation and decision support tools — are probabilistic systems that occasionally generate plausible-sounding but factually incorrect content. In clinical contexts, this means notes that include details not discussed in the consultation, diagnoses not mentioned, or medication references that are inaccurate.

All major platforms acknowledge this limitation and recommend that practitioners review every generated note before signing. This is not optional guidance — it is a clinical and legal requirement. Several platforms have introduced hybrid human-AI review (notably Nuance DAX and DeepScribe) to reduce hallucination risk in enterprise environments, but no platform eliminates it entirely.

Data Privacy, GDPR and HIPAA Considerations

For practitioners in Ireland and across the EU, GDPR compliance is non-negotiable when deploying any AI tool that processes patient data. The EU AI Act (effective 2024–2025) classifies most clinical AI as high-risk AI systems, imposing additional obligations around transparency, logging, and post-market monitoring.⁷

Practitioners should verify the following before adopting any AI clinical tool: whether the platform is GDPR-compliant and where data is stored (EU data residency should be confirmed explicitly); whether patient audio or transcripts are retained and for how long; what the platform's breach notification policy is; and whether patients have been meaningfully informed that AI is being used in their care.

Several platforms — including Nuance DAX — do not store audio recordings, retaining transcripts only. Freed AI stores no patient recordings. Heidi Health and others vary in their retention policies. These should be confirmed directly with each vendor and documented in practice privacy policies.

Cost Realities and Return on Consideration

The cost range is wide. Enterprise systems (Nuance DAX, Abridge, Ambience) run to €500–800 per clinician per month, with multi-year contracts and IT implementation requirements. Mid-tier platforms (Suki, Nabla) fall in the €200–400 range. Accessible entry-level tools (Freed AI, Heidi Health) begin at €90–150 per month, with free tiers available.

For a solo functional medicine practitioner conducting five to eight consultations per day, the calculation is relatively straightforward. Independent analyses suggest practitioners save one to two hours daily on documentation. At a conservative consultation rate, this represents either two to four additional appointments per day — or equivalent time returned to clinical thinking, CPD, and patient communication. At the lower end of pricing, the return is rapid for any practitioner with reasonable consultation volume.

The calculation changes for practitioners with lower consultation volume or those whose notes are brief. In those contexts, the ROI is less compelling, and the priority should be ensuring whatever AI is used is strictly GDPR-compliant and does not introduce new clinical risk through hallucination.

What to Expect in 2026 and Beyond

The Stanford-Harvard report anticipates that AI tools will increasingly move beyond documentation into active clinical decision support — not as replacement for clinical reasoning, but as an additional layer of pattern recognition across complex datasets. For functional medicine, this is potentially significant: practitioners routinely manage patients with multiple interacting systems, complex timelines, and large volumes of laboratory data that are genuinely difficult to synthesise unaided.

Multimodal AI — systems capable of integrating text, genomic data, imaging, wearable data, and patient-reported outcomes simultaneously — is in development across several platforms and research centres. The timeline for clinical validation and regulatory approval is uncertain, but the direction of travel is clear. Functional medicine, with its inherently integrative model, may ultimately be one of the fields best placed to benefit from mature clinical AI.

The practitioners who will navigate this transition most effectively are those who engage with the tools critically, understand their limitations as clearly as their capabilities, and maintain clinical judgment as the non-delegable core of their practice.